This commit is contained in:
jakubknczny 2022-01-22 00:01:03 +01:00
parent 6c296fd9fc
commit 738d542f61
7 changed files with 1 additions and 12077 deletions

2
.gitignore vendored
View File

@ -1,3 +1 @@
mt-summit-corpora
.idea .idea
kompendium_lem*

136
inject.py
View File

@ -1,136 +0,0 @@
import copy
import pandas as pd
import spacy
from spaczz.matcher import FuzzyMatcher
# spacy.require_gpu()
spacy_nlp_en = spacy.load('en_core_web_sm')
spacy_nlp_pl = spacy.load('pl_core_news_sm')
print('lemmatizing glossary')
glossary = pd.read_csv('glossary.tsv', sep='\t', header=None, names=['source', 'result'])
source_lemmatized = []
for word in glossary['source']:
temp = []
for token in spacy_nlp_en(word):
temp.append(token.lemma_)
source_lemmatized.append(' '.join(temp).replace(' - ', '-').replace(' ', '').replace(' / ', '/').replace(' ( ', '(').replace(' ) ', ')'))
result_lemmatized = []
for word in glossary['result']:
temp = []
for token in spacy_nlp_pl(word):
temp.append(token.lemma_)
result_lemmatized.append(' '.join(temp).replace(' - ', '-').replace(' ', '').replace(' / ', '/').replace(' ( ', '(').replace(' ) ', ')'))
glossary['source_lem'] = source_lemmatized
glossary['result_lem'] = result_lemmatized
glossary = glossary[['source', 'source_lem', 'result', 'result_lem']]
glossary.set_index('source_lem')
glossary.to_csv('glossary_lem.tsv', sep='\t')
dev_path = 'dev-0/'
print('lemmatizing corpus ' + dev_path)
skip_chars = ''',./!?'''
with open(dev_path + 'in.tsv', 'r') as file:
file_lemmatized = []
for line in file:
temp = []
for token in spacy_nlp_en(line):
temp.append(token.lemma_)
file_lemmatized.append(' '.join([x for x in temp if x not in skip_chars])
.replace(' - ', '-').replace(' ', '').replace(' / ', '/').replace(' ( ', '(').replace(' ) ', ')'))
with open(dev_path + 'expected.tsv', 'r') as file:
file_pl_lemmatized = []
for line in file:
temp = []
for token in spacy_nlp_pl(line):
temp.append(token.lemma_)
file_pl_lemmatized.append(' '.join([x for x in temp if x not in skip_chars])
.replace(' - ', '-').replace(' ', '').replace(' / ', '/').replace(' ( ', '(').replace(' ) ', ')'))
# glossary
glossary = pd.read_csv('glossary_lem.tsv', sep='\t', header=0, index_col=0)
train_glossary = glossary.iloc[[x for x in range(len(glossary)) if x % 6 != 0]]
# add rules to English matcher
nlp = spacy.blank("en")
matcher = FuzzyMatcher(nlp.vocab)
for word in train_glossary['source_lem']:
matcher.add(word, [nlp(word)])
# add rules to Polish matcher
nlp_pl = spacy.blank("pl")
matcher_pl = FuzzyMatcher(nlp_pl.vocab)
for word, word_id in zip(train_glossary['result_lem'], train_glossary['source_lem']):
matcher_pl.add(word, [nlp_pl(word)])
en = []
translation_line_counts = []
for line_id in range(len(file_lemmatized)):
if line_id % 100 == 0:
print('injecting glossary: ' + str(line_id) + "/" + str(len(file_lemmatized)), end='\r')
doc = nlp(file_lemmatized[line_id])
matches = matcher(doc)
line_counter = 0
for match_id, start, end, ratio in matches:
if ratio > 90:
doc_pl = nlp_pl(file_pl_lemmatized[line_id])
matches_pl = matcher_pl(doc_pl)
for match_id_pl, start_pl, end_pl, ratio_pl in matches_pl:
if match_id_pl == glossary[glossary['source_lem'] == match_id].values[0][3]:
line_counter += 1
en.append(''.join(doc[:end].text + ' ' + train_glossary.loc[lambda df: df['source_lem'] == match_id]['result'].astype(str).values.flatten() + ' ' + doc[end:].text))
if line_counter == 0:
line_counter = 1
en.append(file_lemmatized[line_id])
translation_line_counts.append(line_counter)
print('saving files')
tlcs = copy.deepcopy(translation_line_counts)
translations = pd.read_csv(dev_path + 'expected.tsv', sep='\t', header=None, names=['text'])
translations['id'] = [x for x in range(len(translations))]
ctr = 0
sentence = ''
with open(dev_path + 'in.tsv.injected.crossvalidated', 'w') as file_en:
with open(dev_path + 'expected.tsv.injected.crossvalidated', 'w') as file_pl:
for i in range(len(en)):
if i > 0:
if en[i-1] != en[i]:
if ctr == 0:
sentence = translations.iloc[0]
translations.drop(sentence['id'], inplace=True)
sentence = sentence['text']
try:
ctr = tlcs.pop(0)
except:
pass
file_en.write(en[i])
file_pl.write(sentence + '\n')
ctr = ctr - 1
else:
try:
ctr = tlcs.pop(0) - 1
except:
pass
sentence = translations.iloc[0]
translations.drop(sentence['id'], inplace=True)
sentence = sentence['text']
file_en.write(en[i])
file_pl.write(sentence + '\n')

View File

@ -1,130 +0,0 @@
import spacy
import copy
import pandas as pd
import rapidfuzz
from rapidfuzz.fuzz import partial_ratio
import time
from rapidfuzz.utils import default_process
import sys
spacy.require_gpu()
spacy_nlp_en = spacy.load('en_core_web_sm')
spacy_nlp_pl = spacy.load("pl_core_news_sm")
def read_arguments():
try:
corpus_path, glossary_path = sys.argv
return corpus_path, glossary_path
except:
print("ERROR: Wrong argument amount.")
sys.exit(1)
glossary = pd.read_csv('mt-summit-corpora/glossary.tsv', sep='\t', header=None, names=['source', 'result'])
source_lemmatized = []
for word in glossary['source']:
temp = []
for token in spacy_nlp_en(word):
temp.append(token.lemma_)
source_lemmatized.append(' '.join(temp).replace(' - ', '-').replace(' ', '').replace(' / ', '/').replace(' ( ', '(').replace(' ) ', ')'))
result_lemmatized = []
for word in glossary['result']:
temp = []
for token in spacy_nlp_pl(word):
temp.append(token.lemma_)
result_lemmatized.append(' '.join(temp).replace(' - ', '-').replace(' ', '').replace(' / ', '/').replace(' ( ', '(').replace(' ) ', ')'))
glossary['source_lem'] = source_lemmatized
glossary['result_lem'] = result_lemmatized
glossary = glossary[['source', 'source_lem', 'result', 'result_lem']]
glossary.to_csv('kompendium_lem.tsv', sep='\t')
corpus_path = 'mt-summit-corpora/train/'
skip_chars = ''',./!?'''
with open(corpus_path + 'in.tsv', 'r') as file:
file_lemmatized = []
for line in file:
if len(file_lemmatized) % 10000 == 0:
print(len(file_lemmatized), end='\r')
temp = []
for token in spacy_nlp_en(line):
temp.append(token.lemma_)
file_lemmatized.append(' '.join([x for x in temp if x not in skip_chars]).replace(' - ', '-').replace(' ', '').replace(' / ', '/').replace(' ( ', '(').replace(' ) ', ')'))
with open(corpus_path + 'expected.tsv', 'r') as file:
file_pl_lemmatized = []
for line in file:
if len(file_pl_lemmatized) % 10000 == 0:
print(len(file_lemmatized), end='\r')
temp = []
for token in spacy_nlp_pl(line):
temp.append(token.lemma_)
file_pl_lemmatized.append(' '.join([x for x in temp if x not in skip_chars]).replace(' - ', '-').replace(' ', '').replace(' / ', '/').replace(' ( ', '(').replace(' ) ', ')'))
THRESHOLD = 88
def is_injectable(sentence_pl, sequence):
sen = sentence_pl.split()
window_size = len(sequence.split())
maxx = 0
for i in range(len(sen) - window_size):
current = rapidfuzz.fuzz.partial_ratio(' '.join(sen[i:i + window_size]), sequence)
if current > maxx:
maxx = current
return maxx
def inject(sentence, sequence):
sen = sentence.split()
window_size = len(sequence.split())
maxx = 0
maxxi = 0
for i in range(len(sen) - window_size):
current = rapidfuzz.fuzz.partial_ratio(' '.join(sen[i:i + window_size]), sequence)
if current > maxx:
maxx = current
maxxi = i
return ' '.join(sen[:maxxi + window_size]) + ' ' \
+ glossary.loc[lambda df: df['source_lem'] == sequence]['result'].astype(str).values.flatten() \
+ ' ' + ' '.join(sen[maxxi + window_size:])
glossary = pd.read_csv('../kompendium_lem_cleaned.tsv', sep='\t', header=0, index_col=0)
glossary['source_lem'] = [default_process(x) for x in glossary['source_lem']]
start_time = time.time_ns()
en = []
translation_line_counts = []
for line, line_pl in zip(file_lemmatized, file_pl_lemmatized):
if len(translation_line_counts) % 50000 == 0:
print(str(len(translation_line_counts)) + '/' + str(len(file_lemmatized), end='\r'))
line = default_process(line)
line_pl = default_process(line_pl)
matchez = rapidfuzz.process.extract(query=line, choices=glossary['source_lem'], limit=5, score_cutoff=THRESHOLD, scorer=partial_ratio)
translation_line_counts.append(len(matchez))
for match in matchez:
# if is_injectable(line_pl, match[0]):
en.append(inject(line, match[0])[0])
stop = time.time_ns()
timex = (stop - start_time) / 1000000000
print(timex)
tlcs = copy.deepcopy(translation_line_counts)
translations = pd.read_csv(corpus_path + 'expected.tsv', sep='\t', header=None, names=['text'])
with open(corpus_path + 'extected.tsv.injected.crossvalidated.pl', 'w') as file_pl:
for line, translation_line_ct in zip(translations, tlcs):
for i in range(translation_line_ct):
file_pl.write(line)
with open(corpus_path + 'in.tsv.injected.crossvalidated.en', 'w') as file_en:
for e in en:
file_en.write(e + '\n')

View File

@ -1,30 +0,0 @@
first iteration:
./marian/build/marian --model mt.npz \
--type transformer --overwrite \
--train-sets mt-summit-corpora/mt-summit-corpora/dev/dev.en \
mt-summit-corpora/mt-summit-corpora/dev/dev.pl \
--disp-freq 1000 \
--save-freq 1000 \
--optimizer adam \
--lr-report
next iterations:
./marian/build/marian --model mt.npz \
--type transformer --overwrite \
--train-sets mt-summit-corpora/mt-summit-corpora/dev/dev.en \
mt-summit-corpora/mt-summit-corpora/dev/dev.pl \
--disp-freq 1000 \
--save-freq 1000 \
--optimizer adam \
--lr-report \
--pretrained-model mt.npz
./marian/build/marian --model mt.npz \
--type transformer --overwrite \
--train-sets mt-summit-corpora/mt-summit-corpora/train/train.en \
mt-summit-corpora/mt-summit-corpora/train/train.pl \
--disp-freq 1000 \
--save-freq 10000 \
--optimizer adam \
--lr-report \
--pretrained-model mt.npz

View File

@ -1,12 +0,0 @@
#!/bin.bash
apt install python3-pip
apt install python3-virtualenv
virtualenv -p python3.8 gpu
source gpu/bin/activate
pip install pandas ipython
pip install spacy[cuda114]
python -m spacy download en_core_web_sm
python -m spacy download pl_core_news_sm
pip install spaczz
pip install rapidfuzz

File diff suppressed because it is too large Load Diff

26
test.py
View File

@ -1,26 +0,0 @@
import time
import nltk
from nltk.stem import WordNetLemmatizer
# nltk.download('omw-1.4')
# nltk.download('punkt')
nltk.download('wordnet')
wl = WordNetLemmatizer()
start_time = time.time_ns()
filex = []
with open('mt-summit-corpora/train/in.tsv', 'r') as file:
for line in file:
if len(filex) % 50000 == 0:
print(len(filex), end='\r')
line = nltk.word_tokenize(line)
filex.append(' '.join([wl.lemmatize(x) for x in line]))
stop = time.time_ns()
timex = (stop - start_time) / 1000000000
print(timex)
f = open('temp', 'w')
for line in filex:
f.write(line + '\n')